When an Algorithm Helps Decide Who Loses a Job in California

Yes. A California employer may generally use AI to recommend, select, or help carry out a termination. AI involvement does not automatically make the firing illegal. The key questions are why the worker was fired, how the automated output influenced the decision, and whether the process violated discrimination, retaliation, leave, accommodation, contract, or another applicable employment protection.

Reviewed August 31, 2026. This guide is informational only—not legal, HR, or employment advice. California requirements change, and the rules that apply depend on the employer, worker, system, and facts.

The short answer: AI can assist a firing, but it cannot excuse an unlawful one

California employment-discrimination rules apply to automated systems used to make or assist with employment decisions, including termination. Legal analyses report that the relevant regulations took effect on October 1, 2025 and cover technology that facilitates human decision-making, not only systems that act independently. Using AI in a discriminatory way may therefore be unlawful even when a manager formally approves the result (Norton Rose Fulbright).

Four questions that sound similar can have different answers:

1. Was AI used? Software may have scored, ranked, flagged, or recommended action.
2. Was its output inaccurate? The underlying data, formula, or resulting score may have been wrong.
3. Was the process unfair? The worker may have had no meaningful opportunity to correct an error or explain relevant context.
4. Was the termination legally prohibited? The decision may have involved discrimination, retaliation, protected leave, failure to accommodate, breach of contract, or another unlawful reason.

Evidence supporting the first three points can help investigate the fourth, but it does not automatically prove wrongful termination. An opaque productivity score could be inaccurate or unfair without connecting the firing to a legally prohibited reason.

Employment status also matters. At-will employment does not override anti-discrimination and retaliation protections.

What counts as an automated employment decision

An automated-decision system is a computational process that makes or facilitates a human decision concerning an employment benefit. It may use AI, machine learning, algorithms, statistics, or other data-processing techniques. The Civil Rights Council’s published regulatory text includes systems that screen, evaluate, categorize, or recommend applicants or employees.

The system does not have to issue the termination notice itself. It might generate a productivity score, identify employees below a threshold, rank workers for a reduction in force, flag attendance patterns, or recommend discipline.

Possible inputs include:

• Task-completion data
• Active or idle time
• Response times
• Attendance records
• Performance rankings
• Data about employees obtained from third parties

Ordinary software that merely stores information is different from software that processes information to make or facilitate an employment decision.

How the system is used	Practical question	Useful review focus
AI supplies one input	How much did the output influence the result?	Compare it with reliable, job-relevant evidence.
Employer primarily relies on AI but conducts human review	Did the reviewer check the underlying facts?	Test data accuracy, consider context, and confirm the reviewer can change the result.
Employer relies solely on the automated result	Could anyone inspect and reject it?	Determine whether errors or improper criteria passed directly into the decision.

A manager clicking “approve” does not necessarily transform an automated recommendation into independent judgment. Because the framework covers systems that facilitate human decisions, the relevant issue is what the manager actually reviewed and how much weight the system received.

As a risk-reduction practice—not a complete legal test—HR should ensure that reviewers understand what a score measures, verify that the data belongs to the correct worker and period, investigate discrepancies, consider relevant leave or accommodations, and have genuine authority to reject the recommendation.

When an AI-assisted firing may cross the legal line

The dividing line is not simply whether automation produced an undesirable result. It is whether the system contributed to a result prohibited by employment law.

Disparate treatment generally means that a protected characteristic was intentionally used as a reason for an adverse decision. Disparate impact can arise when a facially neutral rule or metric disproportionately harms a protected group, even without discriminatory intent.

Protected characteristics under California employment law include race, national origin, sex, pregnancy, gender identity, sexual orientation, religion, disability, and age 40 or older, among others. Employer coverage, causation, and the circumstances of the decision remain important.

Hypothetical: disability-related downtime

A productivity system records an employee’s accommodation-related breaks as idle time. It generates a low score, and a manager terminates the employee without investigating the downtime or considering the existing accommodation.

The potential legal issue is not that software measured activity. It is whether the employer discriminated based on disability, failed to account for a required accommodation, or treated protected circumstances as poor performance.

Timing and context may also justify closer review. Discipline imposed soon after an accommodation request, protected complaint, or protected leave can be relevant, especially if the employer’s explanation changes or comparable employees were treated differently. Timing alone does not prove retaliation.

Human involvement may not correct a defective process when the reviewer accepts inaccurate data or problematic criteria. If a model counts approved leave as absence, an approval click leaves the underlying problem intact.

Anti-bias testing can provide useful evidence, but it is not a compliance certificate or guaranteed defense. Its significance depends on the testing’s quality, scope, recency, results, and the employer’s response.

Notice, scores, explanations, and vendor responsibility

The authorities reviewed do not establish a universal California right for every worker to receive notice that AI influenced a termination, an explanation of the system, an opt-out, a score, or access to every related record. Other laws, contracts, policies, litigation procedures, or fact-specific rights may apply, so neither workers nor employers should assume the answer is always the same.

Record retention and employee access are separate questions. An employer may be required to preserve certain employment or automated-system records without every employee having an automatic right to inspect all retained material.

A worker can still make a focused written request:

Please provide the stated reason for my termination, the performance or productivity criteria applied, and any score, ranking, or standard identified as supporting the decision. Please also confirm whether an automated system contributed to evaluating or recommending the action.

Keep the request neutral and specific. Depending on the circumstances, useful information might include the evaluation period, data sources, threshold, comparison group, error-correction process, and identity of the human decision-maker. A request does not guarantee that every item must be disclosed.

Using an outside vendor does not automatically shield an employer. California’s regulatory text addresses agents performing functions traditionally controlled by employers, including technology-assisted employment decisions. That does not make either party automatically liable: coverage, agency, control, causation, contractual roles, and each party’s conduct still matter.

For HR teams, vendor assurances are not a substitute for examining the deployed configuration. Ask what the specific system does, which inputs are enabled, how outputs are presented, what testing applies to the intended use, and whether reviewers can inspect and override recommendations.

California’s No Robo Bosses proposals: what is verified and what is not

Legislative status through August 31, 2026

• SB 7: Governor Gavin Newsom vetoed the bill in 2025. Its proposed sole-reliance ban, notices, human-review requirements, and penalties did not take effect as presented.
• SB 947: The Legislature approved the successor proposal and sent it to the governor by August 31, 2026.
• Practical effect: Legislative passage alone did not make SB 947 effective law as of that status point.

SB 7 would have restricted sole reliance on automated systems for discipline and termination while imposing related review and notice provisions. Because the governor vetoed it, those proposed requirements should not be presented as current duties. The updated SB 7 analysis records the veto.

SB 947 was introduced on February 2, 2026 as a narrower successor. As introduced, it would have restricted sole automated reliance in discipline and termination and included other safeguards. Commentary on that introduced version also described a private right of action and a $500-per-violation penalty, but later amendments may have changed those provisions. Crowell & Moring’s SB 947 analysis expressly addresses the bill as introduced, so those early enforcement terms should not be attributed to the version ultimately approved by lawmakers.

On August 31, Senator Jerry McNerney’s office announced that the Legislature had approved SB 947 and sent it to the governor. The announcement described the measure as prohibiting discipline or termination based solely on an automated system and requiring human oversight, verification, and worker notice when such systems assist. Those were descriptions of a proposed measure—not current duties established merely by legislative passage (California Senate announcement).

Anyone relying on SB 947 after August 31, 2026 should check the current official California bill history, enrolled text, and gubernatorial action before treating any provision as law.

What to do after a suspected AI-assisted firing

Act promptly, but preserve only information you may lawfully access. The immediate task is to determine what reason was given, what data or score was used, who reviewed it, and whether the automated output materially influenced the outcome.

Worker checklist

• Write a dated timeline covering performance discussions, leave, accommodation requests, complaints, discipline, and termination.
• Save lawfully accessible termination letters, write-ups, performance reports, productivity scores, rankings, AI-related notices, emails, texts, screenshots, and applicable policies.
• Preserve accommodation and leave records, plus names and contact information for possible witnesses.
• Record the exact explanation given for the decision and any later changes to that explanation.
• Send a neutral written request for the termination reason, criteria applied, relevant score or productivity standard, and whether an automated system contributed.
• Note potentially comparable situations without taking coworkers’ private records or making unsupported assumptions about protected characteristics.
• Do not enter restricted systems or copy confidential, privileged, proprietary, security-sensitive, or personal information without authorization.
• Consider qualified legal advice promptly because different claims can have different deadlines.

California’s Civil Rights Department may investigate qualifying discrimination, harassment, and retaliation complaints. Submitting an intake form starts the intake process; it does not mean CRD has found a violation. In employment cases, CRD says the intake form must generally be submitted within three years of the last harm, but that is not the deadline for every possible claim. A worker generally must obtain a Right-to-Sue notice before filing their own FEHA employment lawsuit (CRD complaint process).

HR checklist

• Identify the exact system, version, configuration, inputs, outputs, thresholds, and intended use.
• Determine how much the output influenced the proposed termination.
• Verify that source data is accurate, complete, current, and assigned to the correct employee.
• Confirm that the metric is relevant to the actual job rather than merely easy to measure.
• Review accommodations, protected leave, complaints, and other context that may explain the data.
• Examine whether the criterion or result may disproportionately affect a protected group.
• Give the reviewer enough information, time, independence, and authority to reject the recommendation.
• Document what the reviewer checked and why the final decision followed—or departed from—the automated output.
• Clarify employer and vendor roles in data collection, system operation, testing, error correction, and record control.
• Preserve relevant records under applicable retention rules and litigation holds.
• Obtain qualified counsel for high-risk or fact-specific decisions.

Preservation is not proof. A saved score establishes that the score existed, not that it caused the termination. The legally important question is how the system’s data, ranking, or recommendation affected the employer’s decision. Workers should preserve lawful evidence and act promptly if they suspect discrimination or retaliation; HR teams should independently verify automated recommendations rather than rubber-stamp them.